Can long-context AI find all contradictions in scientific literature?
A post says not yet. It describes “high-complexity” tasks whose difficulty grows quadratically with the size of the literature, rather than linearly.
TLDR
The author says finding every contradiction in scientific literature remains beyond long-context AI. They describe a class of tasks whose difficulty scales quadratically with corpus size and say this reverses common long-context model design choices, including block-sparse attention and hybrid models.
Combined views
30.3K
5 Sources, first seen 8h ago
